Unverified50% confidenceOpinionExact time
Existing gradient-based unlearning methods account only partially for the geometric relationships characterized by the leverage matrices, causing them to deviate significantly from the retrained model in output-space behavior.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe core question posed by this study is: when a gradient-based unlearner differs from a model retrained only on the retained data, what structure does this output-space gap have.79% similarUnverifiedThese two leverage matrices are shown both theoretically and experimentally to drive the influence of the unlearning residual on forget outputs and retain outputs.77% similarUnverifiedExisting long-context benchmarks focus almost entirely on recall capability, neglecting state tracking and pattern learning.73% similarUnverifiedOver-reliance on the fine-tuned part during fine-tuning may be precisely the root cause of rare-class capabilities being obscured; directing the model's computation moderately toward the pretrained components may be a more robust fine-tuning route.73% similarUnverified近似遗忘通过梯度上升等方法主动削弱特定数据对模型参数的影响，但效果难以量化验证72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1113846API
curl https://kongchang.com/api/v1/knowledge/claims/1113846MCP
get_claim(id=1113846)